Accuracy

Our report card

Every FPL tool claims to be AI-powered. None of them will tell you how often they are right. This page is our attempt at the opposite: each recommendation is written down before the deadline, scored once the gameweek finishes, and published here whether it flatters us or not.

Live record

Live2026/27· 0 gameweeks scored

Nothing scored yet

Logging starts at gameweek 1. Every recommendation is written down before the deadline and scored once the gameweek finishes.

This section cannot be backfilled. A record built after the fact is not a record, so the only honest thing to show today is how empty it is.

Offline backtest

Backtest2025/26· 38 gameweeks scored

This is a backtest, not a record. It replays 2025/26 offline with the model as it stands today. It is useful for spotting bias and miscalibration, and it is not evidence of live performance - nothing here was published before a deadline.

Captain hit rate34%Our pick finished top 3 of the shortlist we published

Captain regret5.4 ptsPoints behind the best captain in hindsight, doubled

Transfer hit rate79%Suggested transfer ended up net positive

Transfer optimism+1.08 ptsPredicted gain minus what actually landed

Differential hit rate42%Returned 6+ points

xP error1.77 ptsMean absolute error per player prediction

Skill vs naive27.9%Error reduction against repeating last gameweek's points

Worst calibration gap9.0 ppLargest predicted-minus-observed miss across bands

Calibration

When we say 70%, does it happen 70% of the time?

Dot area is the number of observations behind the band. Points above the dashed line happened more often than we said; below it, less. 1 band withheld for small samples (90-100%, n=12).

Where it went wrong

The two ends of the range, so the averages above have something to sit against.

Worst captain weekGW13 · Minteh5 pts, best available was 8 (Rogers)rank 10 of 10

Best captain weekGW31 · Marmoush21 pts, best available was 22 (Senesi)rank 2 of 10

Every gameweek· Offline backtest

The full log, losing weeks included. Selective disclosure would make the whole page worthless.

Every scored gameweek for Offline backtest 2025/26: our captain pick and its rank, the best captain available in hindsight, transfer gain predicted against actual, differential points, and mean absolute error of that week's expected-points predictions.
GW	Our captain	xP	Actual	Rank	Best available	Transfer pred	Transfer actual	Diff pts	xP error	Note
1	Martinelli	6.0	10	9/10	Tavernier (14)	+2.0	+1.9	8	1.94	Captain miss
2	Barnes	6.0	10	1/10	Johnson (11)	+1.7	-0.4	10	1.61	Transfer cost points
3	Saka	4.4	8	1/10	Anderson (9)	+2.6	+0.8	1	2.09	Differential blank
4	Rayan	5.8	16	2/10	Minteh (17)	+1.7	+0.3	1	1.37	Differential blank
5	Watkins	6.0	7	5/10	Osula (10)	+2.5	-0.3	2	1.85	Transfer cost points
6	Gravenberch	5.8	15	8/10	Calvert-Lewin (20)	+2.6	+0.7	3	1.41	Captain miss
7	Estêvão	6.7	15	10/10	Neto (21)	+2.5	+1.6	9	2.00	Captain miss
8	Cunha	4.6	16	1/10	Scott (17)	+0.9	-0.8	2	1.53	Transfer cost points
9	Stach	7.8	18	3/10	Garnacho (20)	+2.9	+1.0	4	1.97
10	Buendía	7.3	19	1/10	Grealish (20)	+2.2	+3.1	1	1.90	Differential blank
11	Tavernier	7.6	13	7/10	Woltemade (16)	+1.9	+2.7	11	1.72	Captain miss
12	O'Reilly	5.3	6	10/10	Hinshelwood (9)	+1.0	-0.9	1	1.61	Captain miss
13	Minteh	8.3	5	10/10	Rogers (8)	+1.8	+1.2	2	2.12	Captain miss
14	Gibbs-White	8.1	13	6/10	Garnacho (17)	+2.1	+1.4	2	1.44	Captain miss
15	Eze	6.6	16	1/10	Hinshelwood (18)	+1.3	-0.2	8	1.46	Transfer cost points
16	Cherki	7.0	17	4/10	Ndiaye (20)	+1.0	-0.5	2	1.54	Transfer cost points
17	Virgil	4.9	10	7/10	Rashford (13)	+1.2	+0.8	4	1.59	Captain miss
18	Estêvão	5.8	9	6/10	Brobbey (11)	+0.9	+0.4	8	2.17	Captain miss
19	Strand Larsen	6.7	7	5/10	Neto (9)	+3.0	+0.7	1	2.10	Differential blank
20	Rogers	7.8	15	10/10	Gakpo (20)	+1.2	+0.2	2	1.49	Captain miss
21	Ndiaye	5.4	9	7/10	J.Murphy (13)	+1.6	+1.5	7	1.58	Captain miss
22	Gibbs-White	5.1	18	4/10	Igor Jesus (20)	+2.9	+0.2	17	1.62
23	Mateta	4.4	13	3/10	Wilson (15)	+2.5	+0.9	3	2.01
24	Tarkowski	7.8	12	8/10	Gakpo (17)	+2.0	+1.3	7	2.15	Captain miss
25	Semenyo	5.3	9	1/10	Richarlison (10)	+2.3	+1.6	2	1.70	Differential blank
26	Virgil	6.9	10	6/10	O.Dango (13)	+2.9	+0.8	0	1.77	Captain miss
27	Bruno G.	6.8	14	9/10	Marmoush (19)	+1.9	+1.0	8	1.73	Captain miss
28	Osula	8.5	7	3/10	B.Fernandes (8)	+1.7	+1.1	16	1.94
29	Richarlison	5.6	17	5/10	Doku (20)	+2.8	+0.7	0	1.71	Differential blank
30	Dorgu	5.4	15	7/10	Cherki (19)	+1.8	-0.1	13	2.09	Captain miss
31	Marmoush	7.4	21	2/10	Senesi (22)	+1.3	0.0	11	2.15
32	João Pedro	4.2	14	4/10	Cherki (16)	+2.2	+1.4	11	1.79
33	Anderson	5.3	17	1/10	Foden (18)	+2.7	+2.7	14	1.44
34	Strand Larsen	4.4	7	1/10	Wilson (8)	+1.3	+0.8	2	2.04	Differential blank
35	Šeško	8.1	6	6/10	Welbeck (8)	+0.8	+0.6	5	1.57	Captain miss
36	Ngumoha	7.5	16	4/10	J.Murphy (19)	+2.4	+2.3	9	1.43
37	O'Reilly	4.8	9	7/10	Thiago (13)	+2.4	+0.9	1	1.69	Captain miss
38	Garnacho	4.3	18	6/10	Hudson-Odoi (22)	+1.7	+1.8	4	2.01	Captain miss

How this is scored

Captain hit: Our pick finished in the top three of the ten-name shortlist we published that week. Not "top three in the game" - only against options we actually put our name to.
Captain regret: Points behind the best captain available in hindsight, doubled because the armband doubles. This is the number that hurts, so it stays on the page.
Skill vs naive: How much our expected-points error beats simply predicting a repeat of last gameweek. An error of 1.8 points means nothing on its own - only against a baseline anyone could build in a spreadsheet.
Withholding: Headline rates need 8 scored gameweeks; a calibration band needs 20 observations. Below that we print "withheld" and the shortfall rather than a number that would only mislead.
Determinism: Every figure comes from ordinary code over a timestamped log. A language model writes the words on this site; it never touches the numbers, and it never gets to grade our own homework.
No backfilling: The live log only ever grows forwards. If a week is missing it stays missing, because a record assembled after the results are known is not a record.

Last scored 3 Aug 2026, 05:27.